Developing Rules

Rules in SQLFluff are implemented as classes inheriting from BaseRule. SQLFluff crawls through the parse tree of a SQL file, calling the rule's _eval() function for each segment in the tree. For many rules, this allows the rule code to be really streamlined and only contain the logic for the rule itself, with all the other mechanics abstracted away.

Running Tests

The majority of the test cases for most bundled rules are "yaml test cases", i.e. test cases defined in yaml files. You can find those yaml fixtures on github. While this provides a very simple way to write tests, it can be occasionally tedious to run specific tests.

Within either a tox environment or virtualenv (as described in the contributing.md file), you can either run all of the rule yaml tests with:

pytest test/rules/yaml_test_cases_test.py -vv

...or to just run tests for a specific rule, there are two options for a syntax to select only those tests:

pytest -vv test/rules/ -k RF01

The -k option simply searches for the content of the argument being in the name of the test, which will match any single or combo tests for that rule. By convention, any test cases for a rule should include the code for that rule.

Traversal Options

crawl_behaviour

Some rules are a poor fit for the simple traversal pattern described above. Typical reasons include:

• The rule only looks at a small portion of the file (e.g. the beginning or end).
• The rule needs to traverse the parse tree in a non-standard way.

Each rule must provide a crawl_behaviour field, either directly or through inheritance, which controls how the parse tree is traversed. Most rules use a SegmentSeekerCrawler, which calls _eval() for every segment of the given type(s). A rule that only needs to look at the whole file once can set crawl_behaviour = RootOnlyCrawler() instead, so that _eval() is only called once, with the root segment of the tree. This can be much more efficient, especially on large files. For example, see rules LT13 and LT12, which only look at the beginning or end of the file, respectively. The available crawlers are defined in src/sqlfluff/core/rules/crawlers.py.

_works_on_unparsable

By default, SQLFluff calls _eval() for all segments, even "unparsable" segments, i.e. segments that didn't match the parsing rules in the dialect. This causes issues for some rules. If so, setting _works_on_unparsable to False tells SQLFluff not to call _eval() for unparsable segments and their descendants.

Performance-related Options

These are other options a rule can set to make linting faster.

provide_raw_stack

Some rules use the RuleContext.raw_stack property to access earlier segments in the traversal. This can be useful, but it adds significant overhead to the linting process, so it is not populated by default. A rule that needs it can request it when defining its crawler by passing provide_raw_stack=True, for example crawl_behaviour = SegmentSeekerCrawler({"newline"}, provide_raw_stack=True).

lint_phase

There are two phases of rule running.

1. The main phase is appropriate for most rules. These rules are assumed to interact and potentially cause a cascade of fixes requiring multiple passes. These rules run the runaway_limit number of times (default 10).

2. The post phase is for post-processing rules, not expected to trigger any downstream rules, e.g. capitalization fixes. They are run in a post-processing loop at the end. This loop is identical to the main loop, but is only run 2 times at the end (once to fix, and once again to confirm no remaining issues).

The two phases add complexity, but they also improve performance by allowing SQLFluff to run fewer rules during the main phase, which often runs several times.

NOTE

post rules also run on the first pass of the main phase so that any issues they find will be presented in the list of issues output by sqlfluff fix and sqlfluff lint.